Fetal Brain Ultrasound Analysis: Sub-Region Segmentation of the Cerebellum and Texture Variability in the Thalamus

Abstract

Ultrasound (US) imaging is routinely used during pregnancy because it is affordable, non-invasive and non-ionizing. Different scans are taken during different gestation stages for various purposes. It is scientifically proven that during 18-22 weeks of gestation, the fetal brain undergoes significant structural growth. This provides an opportunity to extract rich information from fetal brain US images taken during this period. Existing studies on fetal US images solely focus on investigating the size/shape characteristics of brain structures to assess their clinical values. However, the possible clinical value of radiomics, which includes texture features, in relation to fetal growth has been failed to examine. Texture-based analysis is important as it can provide a more comprehensive aspect of growth and this cannot be captured by simple size or shape measurement.

The first study in this thesis applied radiomics-based, including deep learning-based and statistical methods, to extract the textures of the thalamus from the fetal brain US images, then investigated the relationship with gestational age. The results indicate that several texture feature characteristics show statistically significant variations across different gestational age groups at birth. Additionally, texture feature characteristics demonstrate a weak to moderate correlation with the gestational age. This study suggests the feasibility of using texture feature as a novel approach to investigate the fetal brain growth, which is beneficial to future investigation of fetal neurodevelopment or gestational age estimation. Another study of this thesis introduces a novel AMFA Network, incorporating a novel adaptive attention gate and multi-fusion mechanism to enhance segmentation accuracy. Results show that the proposed model outperformed other segmentation arts in segmenting the cerebellum from the fetal brain US images and the breast tumor from the external public BUSI dataset compared to other models.